pytorch/xla

Pull Requests Commits

Refactored the execute cached path.

dlibenzi committed 7 years ago

f52e59f2

Minor code refactoring.

dlibenzi committed 7 years ago

e98052e0

Refactor some code into a function.

dlibenzi committed 7 years ago

777297c9

Re-format python file according to Google pyformat.

dlibenzi committed 7 years ago

120d48fa

Now that we route CHECKs to exceptions, care should be taken when running lambdas.

dlibenzi committed 7 years ago

37baa30c

Added more counters to track parallel apply status.

dlibenzi committed 7 years ago

8cdbd76a

In case of multiple tensors providing data, select the older one.

dlibenzi committed 7 years ago

1f5576ac

Minor code refactoring for the loader wrapper batch counting.

dlibenzi committed 7 years ago

81b0836a

Use a context to capture the ApplyPendingGraph() metadata, to allow sped up operations in case the new apply matches the one within the saved context.

dlibenzi committed 7 years ago

f8a74a57

Allow tensors to reference data from other tensors, so that new generations of inputs and gradient can be tracked back (via the unique ID) to the same tensor kind.

dlibenzi committed 7 years ago

53c74b75

No need to re-check for nullptr as GetApplyOrder() already filtered out the tensors which are noth graph backed.

dlibenzi committed 7 years ago

6df02cdd

dlibenzi committed 7 years ago

4b81126e

Generate the XLA computations in parallel.

dlibenzi committed 7 years ago

5d51ddb8

Rename a few tensor API for consistency.

dlibenzi committed 7 years ago

6024040b

Avoid tensors sync on fused train.

dlibenzi committed 7 years ago

3be07368

Optimize compile to avoid adding destruction work which will be done serially, if a computation is cached.

dlibenzi committed 7 years ago

1a2b2aa2

Route TF_CHECK* to C++ exceptions to give a chance to the upper layers to catch.

dlibenzi committed 7 years ago

9c434399

Now that ConfigureDistributedTPU releases all the TPU memory, we do not need to tidy up on exit and we can simplify the code.

dlibenzi committed 7 years ago

9054e5d6

Enable more instant tracking of the samples per second rate.

dlibenzi committed 7 years ago

c7c80106

Update the PyTorch commit id to latest

asuhan committed 7 years ago

fc8a5667

Add client side compitation cache.

dlibenzi committed 7 years ago

01aa6b66

Prepare XRT computations in parallel.

dlibenzi committed 7 years ago

f2bf8c4e

Move InputListAttr to anonymous namespace

asuhan committed 7 years ago

f8d36d3b

Some code restructuring and naming cleanup.

dlibenzi committed 7 years ago

a4b3d278

Fix convolution to fit new arguments.

dlibenzi committed 7 years ago

4df06aa3

Fix XRTAllocate to not use the now-goge API anymore.

dlibenzi committed 7 years ago

e65b99cd

Fetch new TF bits to unbreak CrossReplicaSum HLO proto changes.

dlibenzi committed 7 years ago

113af200

Build upstream pytorch as part of nightly torch_xla builds

jysohn23 committed 7 years ago

Verified 67459323

Build upstream pytorch as part of nightly torch_xla builds

jysohn23 committed 7 years ago

a50bb35a

Fix several warnings

asuhan committed 7 years ago

6bd85fa2

Newer Older